Best Practices Checklist

The checklist below helps you avoid common pitfalls before finalising your study.

Numerical stability

Verify CFL stability and time-step size. Reduce the CFL parameter if you see oscillations, waves, or sudden thickness spikes (a value below 0.5 is typically safe). Also examine dt in the output: very small adaptive time steps can indicate near-instability. See Numerical Tips: Time stepping for details.
Vary solver settings. Check that results do not change significantly when you double nbit (iceflow solver iterations). If using the network emulator with periodic re-training, also test sensitivity to the retrain frequency and learning rate — these three interact.
Validate the iceflow solver. If using mapping: network, run at least one short test with mapping: identity to confirm the emulator is not introducing significant error for your domain. Note that lr / lr_init differ substantially between the two (~0.9 for identity, ~1e-5–1e-3 for network) — see iceflow: Practical guidance for guidance.

Volume and area time series. Plot total ice volume and area over time. Abrupt jumps or monotonic growth to unrealistic values are red flags.
Mass balance. Confirm the chosen SMB method is appropriate for your application — see the FAQ for a quick comparison of the smb methods (simple, oggm, accpdd).
Ice velocity. Compare modelled surface velocities against observations where available. Values exceeding a few km yr⁻¹ for alpine glaciers are unusual.

Cost function convergence. Residuals should decrease monotonically (at least on average). A stalling or erratic cost function often points to a learning rate or regularisation issue.
Regularisation. Apply regularisation (e.g. regu_slidingco, regu_arrhenius) to avoid over-fitting noisy observations — verify that inferred fields are physically smooth.
Inversion step count. A modest number of inversion steps is usually sufficient for a good initialisation; the forward simulation is more sensitive to the initial geometry than to the exact inversion accuracy.

See Numerical Tips for deeper guidance, including scalar parameter calibration with Hydra + Optuna.

Sensitivity to physical parameters. Use Hydra multirun to test key parameters such as the Arrhenius factor, sliding coefficient, or ELA gradient:
```
igm_run +experiment=params \
  processes.smb.simple.array.1.1=0.005,0.007,0.009 \
  --multirun
```
Sensitivity to initial conditions. If initialising from observations via data assimilation, check how sensitive projections are to the initial ice thickness.
Run multiple times. Due to the stochastic nature of neural-network training (random weight initialisation, mini-batch sampling), results can vary between runs. Repeat the simulation at least a few times and verify that key outputs (volume, velocities) are consistent across runs.

Record the configuration. Every run saves its full resolved configuration in .hydra/config.yaml. Keep this file with your results.
Record the IGM version. Log the commit hash printed at startup, or note the version installed via pip (pip show igm).
Use version-controlled params. Keep your params.yaml in a git repository alongside the analysis scripts.